Teaching LLMs to Admit Uncertainty in OCR

Vision language models (VLMs) are increasingly replacing traditional OCR pipelines, but on visually degraded documents they often hallucinate, producing fluent yet incorrect text without signaling uncertainty. This occurs because current post-training emphasizes accuracy, which encourages models to guess even when uncertain. The problem persists in state-of-the-art systems and severely impacts OCR reliability. To improve the trustworthiness of OCR on degraded documents, we propose uncertainty-aware OCR. Rather than suppressing guesses, our model transcribes while explicitly bracketing spans it deems unreliable with uncertainty tags. To train our model, we use Group Relative Policy Optimization (GRPO). We define the usage rules for uncertainty tags and an evaluation protocol. We introduce a pseudo-labeled cold start and a multi-objective reward that balances transcription accuracy and uncertainty coverage while preventing reward hacking. We explore different combinations of cold start and reward granularity and verify the effect of reward parameters in preventing reward hacking and improving the corresponding metrics. We also introduce Blur-OCR, a challenging OCR benchmark for uncertainty-aware OCR on degraded documents. In detailed experiments, our model maintains transcription accuracy while achieving an uncertainty tag F1 score of 0.685, and substantially outperforms both open- and closed-source baselines.